Leela Chess Zero Refutes The Dutch Defense? | Computer Chess Championship Finals

[@chess]

Watch more Computer Chess games here! kzbin.info/aero/PL-qLOQ-OEls7uZ_Kr0zY8AgfYavPtUaH_

[@MrDannyg77]

Love Danny’s enthusiasm and why wouldn’t he be so animated. Leela’s play in this game is the stuff of legend.

[@Neisey]

Danny Gee for some reason Danny looks like he aged ten years in like two weeks. Maybe he lost weight? Or started doing drugs. Maybe he’s just plain old aging

[@derda3209]

stockfish: "thats not a good strategy!" leela: "strategy? chess is only tactics."

[@homerp.hendelbergenheinzel6649]

True. Lol

[@michaelr1661]

Except Leela is better at strategy than Stockfish. You ain't beating Stockfish easily searching 30 or more (half)moves ahead in tactics

[@ryan82scott]

Just a formatting thought: when you are exploring alternate scenarios, have you considered a different coloured board? That might help with differentiating between the actual game and the proposed alternative.

[@danielrensch9238]

RSM82 good thought!

[@raheyna8888]

Is this game from the TCEC going on right now? I can't find this game..

[@sosotik]

And here I am thinking quarantine was supposed to make people like you smarter

[@ryan82scott]

@@sosotik Not sure if I understand the hate... if you think it's a bad idea, go ahead and say so; say why if you can. Certainly not worth the words being directed towards me. It was just a suggestion, not an order or direction. It could easily be ignored or rejected without the garbage.

[@ryan82scott]

@Deep Mind X Why?

[@andrejbogdanov2816]

You really excel at these kinds of videos. Thank you very much!

[@danielrensch]

Thanks man!

[@lousyfication]

Exactly! Excelent it is. Thanks a lot.

[@Arc_Soma2639]

"An abusive Russian chess coach would come along and just sweep those off the board and say something like: You are not using them anyway why you want them? And just throw them away" HAHAHAHA MY F** SIDES

[@solomu8772]

Great line. My sides got destroyed.

[@DJHastingsFeverPitch]

A la Searching for Bobby Fischer

[@ozboltmenegatti]

an actual lol moment this was

[@bububuracing62]

Confirmed, that actually happnd to me :-D

[@ilyakalinin2660]

Да, угарнул с этого)

[@exponentmantissa5598]

You didnt explain the horizon effect correctly. The horizon effect is when the computer calculates the position out to a large number of ply and heads down that tree branch only to discover that just beyond where its calculations ended (the horizon) the game is dead lost. For instance a computer calculates 32ply deep and it looks favorable but at the next ply level the position is losing. Kasparov played a great game against deep blue in his first match (game 6) where the computer headed down a branch of the tree that had all dead losses. After a couple of moves the computer then sees the correct eval but it is too late, it had passed the point of no return.

[@surfingbilly9654]

best explanation ive seen so far

[@Harrs2]

Yeah, but what Danny said wasn't lengthy and got the point across

[@zergreenone8111]

@@Harrs2 It's possible to explain it simply and correctly. The horizon effect is when an engine doesn't look far enough ahead to accurately evaluate how good a certain move is, it only looks ahead up to a certain "horizon".

[@mahyu21]

Intuition?

[@THEBOYDALEK]

zzzzzzzz

[@Nortindevtech]

That is crazy, one of the most beautiful games I have seen recently. Amazing analysis Danny

[@StoicTheGeek]

Danny's analysis of 1. d4 f5 - "maybe it was a mouse slip". 😂😂 Absolute gold.

[@durgeshkulkarni1412]

15:22 "The queen lives for black at the end of variation like this,,unfortunately THE KING DOES NOT" -SIM Daniel Rensch 2020

[@josiahroyer1062]

The operation was successful, but the patient died.

[@DoronBarShay]

Incredible line

[@isehaji3792]

God Save The Queen on that tragic Night at the Opera

[@mediocre2master9]

I just wanted to say thank you, Danny, for always inspiring my chess improvement with your insight, energy, and understanding of the game. I also wanted to thank you for kicking my ass so convincingly in a bullet brawl in 2015 that I really hunkered down and decided to start improving my game; after some breaks due to college I finally crossed the 2000 online blitz threshold last month. You’re an inspiration, man!

[@mediocre2master9]

Oh he 100% would still win every game 😅

[@mileseramosroman2387]

Wow bullet brawls...good times

[@danielrensch9238]

Way to go man!!!

[@yoshiboy6198]

I think Leela "trolls" just to get into simpler positions, Leela knows how to win the queen vs rook endgame and the pawn just complicates things a little bit. Giving up the pawn leads to a simpler game which is still a definite win for Leela.

[@italixgaming915]

That's also what I think. I saw another game by Leela where she gave back material because she saw that she was completely winning even with that non optimal move. For humans, keeping their material advantage is a natural trend but Leela doesn't care, as soon as she sees a forced win line, she goes for it and doesn't bother searching another one that would seem more natural to a human.

[@JG-zs8tr]

Yep. Lc0 has learned the entire game "backwards" starting with the end of the game, so she's always just trying to reach a line she already knows to be winning. There's no need for her to experiment and learn about shorter paths to victory if she already has a known victory in sight. This can lead to moves that look like trolling humans; for example, underpromoting pawns to rooks.

[@austinnunez2031]

I forgot where I read or heard but Leela is doing something like cutting out any possibility for loss. She follows the Monte Carlo tree search and once she finds herself winning, she begins to essentially constrict the opponent, slowly taking away any moves that could lead to her loss or draws, stalemates perpetuals etc. It’s the reason she doesn’t finish games quicker, she essentially “trolls” the opponent with a “there’s nothing you can do”. It really is fascinating. She’ll give pieces back just to misplace your pieces and you have absolutely no counterplay to her plan. Total utter domination. Beautiful.

[@tolkienfan1972]

I think it's more that it doesn't matter. They both win. There's a kind of randomness that comes from the weights in a neural net. It's not like it's an more or less effort for a machine. Effort is meaningless here.

[@tolkienfan1972]

@@bishshoy would you prefer it if I used the term unpredictability, Mr Pedantic?

[@Menaceirl]

Leela "trolling" is a result of neural networks in all games. They find the path of least resistance more often than they find the "best" way to solve a problem. this can be seen in nearly every implementation of neural networks. Leela may see getting more pieces of the board as simplifying the position just as we do and in turn, would be easier for her to convert.

[@lagerbaer]

This is similar to the tongue-in-cheek advice: "If you can take the opponent's queen, OR mate in 5, you should take the queen".

[@homerp.hendelbergenheinzel6649]

removing the pawn with this troll move also god rid of countless possible stalemate options/positions for black i think.

[@unathorizdwatermelon]

Loving the Russian accent danny.

[@danielrensch]

And it loves you!

[@nathansharp3193]

Daniel Rensch, the abusive Russian coach only truly loves the game of chess.

[@1999yasin]

@@nathansharp3193 That description was oddly specific lmao

[@igorm.3114]

As a Russian I had a good laugh there :)

[@tonistaak]

@@danielrensch can you analyse game number 16 where LC0 white lost to Stockfish black

[@AndyChamberlainMusic]

Next time: Leela Chess Zero refutes 1.e4

[@MrZluvu4ever]

Andy Chamberlain Music perhaps you mean 1.d4!?

[@ihateorphans]

@@MrZluvu4ever Probably 1.a3

[@danielrensch9238]

Ha! Don’t tempt her 😉

[@jeffgreen3376]

1. e4 is easily refuted by 1. c5 (the Sicilian Defense). Also, 1. e6 (the French Defense) can equalize.

[@kdhavle]

I think Alpha Zero did that. AZ never played 1.e4, if I remember right. :)

[@ZodiacKillerSFPD]

Great video. Your analysis was as evident as your enthusiasm!

[@danielrensch]

Appreciate that.

[@ph6560]

I haven't seen too many of these instructional videos (involving openings) by Rensch in recent times, but really liked his way of presenting the ideas. However, hope he'll continue make much more of these!

[@adamstone5865]

That picture had me laughing so hard, “one does not simply play f4” so well done and all chess players can appreciate that

[@omarabdelhafiz]

The caption says f5

[@sensorcato]

8:09 1...f4 Qxb2 2.Rb1 Qf6 3. O-O Goosebumps...

[@somtovitus]

I love the Russian chess coach joke. Seemed so real.

[@Nf5]

I wish you said when the opening ended and the engine moves started. @9:20 "that you created Stockfish when you moved your f pawn on move one" I think most people who aren't used to computer chess won't understand that Stockfish didn't decide to play f5.

[@PiercingSight]

I watched this game live and freaked out in the chat with everyone else starting from d6. From d6 it was clear that Leela was just going to crush positionally and not only did she do that, but she also crushed tactically, and taking black's queen was one of the most thrilling moments to watch. We had a party in the chat. Absolutely incredible game! Thank you for the analysis!

[@Patralgan]

This reminds me of AlphaGo playing Go and sometimes instead of grabbing a large advantage, it would instead play something which seems relatively insignificant. This puzzled many humans. The reason was that since you only need to win by 0.5 points to win the game, AlphaGo would go for smaller advantage if it is more confident it would lead to victory instead going for larger advantage, which is not a safer choice (but in the games they would have been incredibly safe nonetheless). A human player would go for a larger advantage to maximize chances to win.

[@elijahbuscho7715]

So like humans would lean more towards thinking about material, ie. you wouldn't want to sacrifice material if you don't have to, because the more material you have, the more likely you are to win, whereas the computer thinks more in terms of positions, and will play in a way to get to a certain position, from which there are fewer other positions in which they lose. Because really, big advantage doesn't matter, all the matters is enough advantage to get the checkmate.

[@Patralgan]

@@elijahbuscho7715 yeah, I guess that's rather accurate

[@Jaycafka]

Great video, thank you! It was not only entertaining but also very educational.

[@rupen42]

8:23 Ben Finegold says to not play f3, not f4. He often uses his "never play f3" adage to illustrate when to play f4. But yeah, f4 is also a weakening move.

[@wayneliang4524]

Because the algorithm would only be changed when lost,the " trolling " moves or the unnecessary moves wouldn't be ruled out as long as it's winning, therefore it still plays it

[@iliyzavialov3199]

So much passion in commentary, amazing)) btw much better then agad does in these days

[@funchable212]

Amazing. Please make more of this type of video Danny you’re the man!

[@Radjehuty]

we really don't know why she "trolls" but it always happens when there are numerous winning lines. If you can imagine that all her knowledge comes from playing against herself, it's possible that in enough games to be significant, perhaps she's managed to stalemate herself which caused her to seek lines that are simplified and might explain why she prefers to sac her pieces for the familiar QvK or RvK endgames. It has also caused her to shuffle in to draws from winning positions though. A future(soon) patch is aimed at solving this hopefully.

[@fixpontt]

not trolling, simplifying into tablebase position

[@ddimin]

We DO know why lczero "trolls", it's not related to tablebases (it does this even without tablebase) and it's not because of reducing stalemate chances. It's not chess-specific, AlphaGo does the same. The reason why it does that is because its end criteria is to maximize "winning probability". Lczero just don't know the difference between two moves if it is sure that they are both winning, so it just chooses randomly.

[@ddimin]

@@danispringer It's because if you ten times toss a coin then 99.9% of the cases you get at least one heads. Same here: lczero has many opportunities to sack material.

[@DJHastingsFeverPitch]

@@fixpontt that may be true, except I don't believe that Leela has tablebases? I could be wrong?

[@fixpontt]

@@DJHastingsFeverPitch go to tcec webpage, you can see after every move that how many tablebase position L0 found in that move, it does not use that as extensively as Stockfish but it uses it nonetheless

[@JV-ge8bm]

Don’t tell Simon Williams!

[@gonzalo4658]

Jared Vu you’re goin to hell for that

[@reidpattis9478]

Come on 'arry.

[@12jswilson]

Leela gives and it takes. It refutes the Dutch, but shows that h4 is almost never a bad move

[@davidcopson5800]

BOOM!

[@justdavelewis]

Shane Wilson I think Simon would be ok if he knew that a gambit and pushing Harry was a Dutch refutation, I think that’s the only situation he’d be ok with

[@johannesb9681]

This is the best game I have seen this year so far. Amazing stuff

[@haykmanukyan5462]

great! wanna see more analyses by Dani Rensch

[@christopherhume1631]

I lived in AZ until May of 1999 and began to get back into Chess during that period. Tal Shaked was a young rising star then. He had a Chess Column, which I believe appeared in the Arizona Republic for a time. I imagine you would have been aware of this growing up. I have been enjoying your humor, analysis and instructive approach for several years now. Always fun to watch you and Hess quarrel.

[@wendestein9092]

thanks for both the good analysis and humor Daniel!

[@plane_riotous4202]

A beautiful game indeed. Daniel’s presentation along with teaching and analysis is spectacular. Plenty of effort and time he’s been dedicated for this video. I really loved it and thank you very much for that. Stay safe, stay sane🤪

[@scarletevans4474]

3:55 Danny says that there should be some link to analysis, but I can't find it anywhere. Did someone forgot to put it in the video description? Anyone knows where I can find it?

[@dannygjk]

ikr I suppose it's at chessdotcom.

[@MrLimon27]

Superb! Thank you for bringing us these amazing games with deep analysis.

[@durgeshkulkarni1412]

Great dedication to stream the video just after finishing an analysis, Danny!! PS- The last part where you mentioned why would any neural network want the game to last longer than it should will keep me driving all day, motivates me to search more on NN and Machine Learning !!! :-D

[@willlacey7621]

Daniel’s commentary is always, without a doubt, top notch. Even throwing that bit about his boys playing table tennis cracked me up.

[@alexnaturalis1179]

I never thought chess commentary could be so entertaining. Killing it Daniel!

[@bhgtree]

Wonderful game and commentary, this has just become one of my alltime favourite Leela games. Thanks Danny.

[@ig2d]

the sac of the queen is because of the Table base. sacking the queen is not the shortest path to mate but it is the shortest path to a winning entry on the tablebase

[@AmorLucisPhotography]

Lovely game. Thanks for explaining it. As a fan of the Dutch (Stonewall) and after getting stung by the Staunton Gambit too many times, I now only enter the Dutch by transposition.

[@maxtsivourakis137]

Thanks for making these videos Danny

[@DuncanEduardo]

Brilliant and intelligent chess commentary and analysis. Great raconteur. Amazing game

[@pantsik2]

Great stuff! I would like to learn more about Leela trolling with sacrificing pieces that lead to longer mate paths. It seems like Leela is prefering fewer pieces on board and longer mates that more pieces and shorter mates.

[@Zen-rl5pv]

Not going to lie the only thing better than these games is your commentary, love it haha

[@younis24de]

8:22 Ben Finegold never disapproved of f4. Pawn to f3 on the other hand…

[@ig2d]

I absolutely love your commentary.. If you're looking for another game to analyse check out tcec game 35. A classic sicilian dragon with all the standard motifs you would come to expect (opposite casting pawn storms on both wings exchange sac on c3 etc...) played briliantly by both entities... At 5:17 it seems like black might just get away with e6(?)

[@lasker500]

incredible play by Leela. So much to discover with this new engine.

[@dg1431]

There's something weird about these games. At 12:49 the engine gives black about a -9 advantage so I decided to put this into my phone and check it out. Stockfish on my phone rates this position as basically equal giving it -0.3 in favor of black. How come the software on my phone understands the position better than this version of stockfish?

[@JoeBruin96]

.Absolutely love the thumbnail. Oh and Danny you’re the best

[@latibes]

Kingcrusher: Stockfish has 'spectator' pieces

[@vecter]

Look at them! They're all spectators!! [in British]

[@nicholasperkins4655]

They are off in Siberia!

[@friedrichroell4018]

Great video Danny, cheers!

[@ehatipo4598]

you. are. crazy. good. Please more tournaments with gingergm etc. we all here to watch! Thank you for the crazy content!!!

[@thorstenjaeger1203]

Great stuff for me!!! I like to play the Staunton and I LOVE(!) Queen vs. Rook, studying that for months(!?!). Thanks a lot!!!

[@sensorcato]

*Bobby Fischer:* e4 is best by te- *Leela:* I just refuted it. *Bobby Fischer:* 😨

[@jeffgreen3376]

I think that the Sicilian Defense refuted 1. e4 long ago. lol

[@Bustical]

The first move (for both colours) were chosen by humans The first move Leela played was 2.e4

[@nicholasperkins4655]

@@jeffgreen3376 Leela's opinion on the Sicilian is an interesting one. She actually thinks the Moscow variation (Bb5+) is white's best option against the d6 Sicilan, however her favorite form of the Sicilian are the e6 variations. Yasser Seirawan would be pleased. Because of how ECO codes work, I am uncertain what Leela prefers as white against the e6 Sicilian. I know Bobby Fischer preferred the KIA.

[@jeffgreen3376]

@@nicholasperkins4655 Interesting. I have more trouble with the Exchange Rossolimo/Moscow Variation than most other tries by white (I play 2. Nc6 and 3. d6 in most Sicilians). I can beat the Exchange Rossolimo (3. Bb5) with a 2. e6 system, but it doesn't work well for me in other (non-Bb5) lines, so I normally stick with an early d6, but often play e6 or g6 a few move later (on moves 4--8) and transpose into a Scheveningan or Dragon Sicilian.

[@Earwaxfire909]

Beautiful game and great analysis. Thank you!

[@chesscomdpruess]

OMG! Danny and Leela just blew my mind a dozen times!

[@heffalump111]

awesome video, I love your enthusiasm

[@alchemist9461]

Great game! Thank you for the coverage! That 19...Rf2 20. Rxg6! Qe4 21. Rxg7! Kxg7 22. Nf4!! just stuned me! Just wonderful!

[@bwatspro]

These are my favourite kind of videos :D Danny the best

[@JimJWalker]

"She plays with her food..." LMAO

[@seasideman]

Amazing game, thanks for the great analysis Danny.

[@EwanMellor]

Great analysis, but that’s totally not why it’s called a “horizon effect”. It’s because you can only see so far ahead before things are hidden below the horizon. No matter how much you understand about what’s close (imminent moves), a depth search like Stockfish’s won’t help with what’s beyond the horizon.

[@visiblethoughts]

Danny is the best commentator ever. I could listen to him talk about about the color brown.

[@johnaustin704]

Great analysis and great enthusiasm = great video!

[@gonzalo4658]

This is the kind of daniel rensch that i like

[@kamvamatsila7576]

I Love This guy! very funny and insightful!

[@vybor]

A great game... and a great presentation! Many thanks.

[@ZorbaPress]

Incredible game -- thanks Leela and Daniel.

[@ceferinojrvizco5832]

Tnx for sharing ..i realy enjoy watching it wit analisis

[@matthewtaylor9906]

Enjoyed it Danny!

[@rhalleballe]

Actually, i played this game with Komodo 13.3, and after move 19. Rg4 Komodo shows up black Rf6 as best move for black, but still with big disadvantage (-1.7) for black! So Komodo did it right! If Rf2 instead, Komodo plays: 20. Rxg6 Qe4, 21. Rg3! Qxe2, 22. Qxe2 Rxe2, 23. Bh6 etc. with total demolition (even without Queens!).

[@rickkarras9708]

Love the screen shot

[@glenncooper3524]

Aren't the first 15 moves openings there forced to play. Plus I looked twice couldn't find this final position or opening

[@xplicitmike]

That is an incredible game. Thank you for the analysis.

[@RKBenchmarker]

Wow, brilliant analysis on a brilliant game!

[@GinjouArmy]

00:58 D-man spitting a few quick bars.

[@mattjensen3499]

Great video Danny!

[@danielrensch9238]

Matt Jensen thanks Matt!

[@JimMalmPHOTO]

What flare. I could listen to this guy all day.

[@poprockssuck87]

If I had to guess, I would say that Leela "trolls" because reinforcement learning has a reward structure. Leela is probably set up so that it rewards incrementally throughout the game rather than just once right at the end for a win. This is because it makes it faster to train the neural network, though it may not lead to the most optimal results (but those most optimal results may not be practically attainable anyway). Therefore, the neural net probably prioritizes winning most of all, but if it can be winning for the longest number of moves, that's better because it receives a bigger reward. That, or Leela is mainly focused on winning independent of the number of moves it takes. It may just focus on the likelihood of winning for a given position rather than winning in the most precise way.

[@perfect_harmony4348]

Wouldnt it shuffle around alot more in winning positions by that logic? If you were right the optimal game strategie would be to get a as many twofold repetitions and as close to 50 moves as possible in each winning position to maximize the move counter

[@poprockssuck87]

@@perfect_harmony4348 Once it finds a continuation that ends the game, it chooses that. The neural network just makes the process of evaluating open-ended positions less contrived because there's very little human manipulation of the evaluation process.

[@dtracers]

@@perfect_harmony4348 it often does. It is something we are actively trying to fix. Because it can lead to her losing winning position because she waits so long

[@vecter]

@@dtracers do you work on Lc0?

[@ugood]

Thank you Danny 💗

[@AnnoShark]

awesome stuff again! thanks for the great analysis. would be really interested how Lela uses the Dutch from the Black side. any analysis coming for one of these games?

[@leavethebasket7435]

Fucking love Danny he's so funny and charismatic, not to mention a fantastic teacher

[@john2001plus]

Endgame tablebases is why Lela played the ending the way she did. She saw that it was forced mate all the way to the end.

[@xzjulius]

Love the commentary

[@ImmortalCreature]

stylish game and great pragmatical explanations

[@clownfishchess4501]

That was an impressive game, but where is the reference? I was looking for the game pgn.

[@twoking10]

All of Stockfish Queen side pieces totally tied up. Leela down material by one minor piece in the count, yet... in one sense Leela is actually ahead in material because Stock has 3 pieces that can't seem to get in the game.

[@Aoltooliol]

Leela doesn't troll. It's just that when there are several winning lines she picks the one where she has to do the least amount of calculation. It's just efficiency.

[@blazingblitzen]

The Mouse slip at 9:23 😂😂

[@Nidhogg86]

Good stuff. Anyone else wanna see Seirawan chess, S-Chess, SHARPER or whatevs with the engines going at it? It would be great to just have it available online somewhere, as the two extra pieces of dynamism should have at least us and probably also the neural networks busy for quite a while with all the new possible complications available.

[@guyscounter]

The sales of GM Simon's Rebooted Killer Dutch will be halted.

[@patzergamer]

Do you recall what year you were in Reno? I’ve been going to the tournament for 8 years, hard to think I missed an opportunity to see Danny!

[@danielrensch9238]

Kevin C I haven’t been in a long time... the Sand me Regency. Weikel’s events...

[@Adalekoth]

The link to the written analysis you mentionned is missing in the description, could you fix that please ?

[@josiahroyer1062]

This is great commentary!

[@AnkhArcRod]

I gave a like without watching the video! Awesome thumbnail!

[@BettyCastella]

I think the reason leela made that bizarre queen sacrifice at the end is that she, like most chess engines, has memorized all endgames involving 6 pieces, and so she knew that if she sacrificed the queen, she could win without having to think, wheras if she pushed the pawns, she'd have to think for a bit longer. You ask "why make the game take longer than it needs to?" but actually for her, it makes the game shorter, in terms of milliseconds on her chess clock.

[@tysak4]

😂😂 great content Danny we want more!